Why did a vehicle stop? A methodology for detection and classification of stops in vehicle trajectories

ABSTRACT

Trajectory data mining is a lively research field in the domain of spatio-temporal data mining. Trajectory pattern mining comprises a set of specific pattern mining methods, which are applied as consecutive steps on a trajectory with the goal to extract and classify re-occurring spatio-temporal patterns. Despite the common nature and frequent usage of such methods by the GIScience community, a methodological approach is missing so far, especially when it comes to the use of machine learning-based classification methods. The current work closes this gap by proposing and evaluating a machine learning-based 3-steps trajectory data mining methodology using the detection and classification of stop points in vehicle trajectories as example. The work describes in detail the applied methodologies with respect to the three mining steps ‘stop detection’, ‘feature extraction’ and ‘classification in traffic-relevant and non-traffic-relevant stops’ and evaluates six machine learning-based classification algorithms using a real-world dataset of 15,498 vehicle trajectories with 5,899 detected stops (thereof 2,032 manually classified). Due to its exemplary nature, the presented methodology is suited to act as blueprint for similar trajectory data mining problems.     

Automated terrain feature identification from remote sensing imagery: a deep learning approach

ABSTRACT

Terrain feature detection is a fundamental task in terrain analysis and landscape scene interpretation. Discovering where a specific feature (i.e. sand dune, crater, etc.) is located and how it evolves over time is essential for understanding landform processes and their impacts on the environment, ecosystem, and human population. Traditional induction-based approaches are challenged by their inefficiency for generalizing diverse and complex terrain features as well as their performance for scalable processing of the massive geospatial data available. This paper presents a new deep learning (DL) approach to support automatic detection of terrain features from remotely sensed images. The novelty of this work lies in: (1) a terrain feature database containing 12,000 remotely sensed images (1,000 original images and 11,000 derived images from data augmentation) that supports data-driven model training and new discovery; (2) a DL-based object detection network empowered by ensemble learning and deep and deeper convolutional neural networks to achieve high-accuracy object detection; and (3) fine-tuning the model’s characteristics and behaviors to identify the best combination of hyperparameters and other network factors. The introduction of DL into geospatial applications is expected to contribute significantly to intelligent terrain analysis, landscape scene interpretation, and the maturation of spatial data science.     

Identification and mapping of riverbed sediment facies in the Columbia River through integration of field observations and numerical simulations

In the Hanford Reach of the Columbia River, a thin layer of recent alluvium overlies the sedimentary formations that comprise the unconfined groundwater aquifer. Experimental and modelling studies have demonstrated that this alluvial layer exerts significant control on the exchange of groundwater and surface water (hydrologic exchange flux), and is associated with elevated levels of biogeochemical activity. This layer is also observed to be strongly heterogeneous, and quantifying the spatial distribution of properties over the range of scales of interest is challenging. Facies are elements of a sediment classification scheme that groups complex geologic materials into a set of discrete classes according to distinguishing features. Facies classifications have been used as a framework for assigning heterogeneous material properties to grid cells of numerical models of flow and reactive transport in subsurface media. The usefulness of such an approach hinges on being able to relate facies to quantitative properties needed for flow and reactive transport modelling, and on being able to map facies over the domain of interest using readily available information. Although aquifer facies have been used in various modelling contexts, application of this concept to riverbed sediments is relatively new. Here, we describe an approach for categorizing and mapping recent alluvial (riverbed) sediments based on the integration of diverse observations with numerical simulations of river hydrodynamics. The facies have distinct distributions of sediment texture that correspond to variations in hydraulic properties, and therefore provide a useful framework for assigning heterogeneous properties in numerical simulations of hydrologic exchange flows and biogeochemical processes.     

An AI-based model for describing cyclic characteristics of granular materials

Modelling cyclic behaviour of granular soils under both drained and undrained conditions with a good performance is still a challenge. This study presents a new way of modelling the cyclic behaviour of granular materials using deep learning. To capture the continuous cyclic behaviour in time dimension, the long short-term memory (LSTM) neural network is adopted, which is characterised by the prediction of sequential data, meaning that it provides a novel means of predicting the continuous behaviour of soils under various loading paths. Synthetic datasets of cyclic loading under drained and undrained conditions generated by an advanced soil constitutive model are first employed to explore an appropriate framework for the LSTM-based model. Then the LSTM-based model is used to estimate the cyclic behaviour of real sands, ie, the Toyoura sand under the undrained condition and the Fontainebleau sand under both undrained and drained conditions. The estimates are compared with actual experimental results, which indicates that the LSTM-based model can simultaneously simulate the cyclic behaviour of sand under both drained and undrained conditions, ie, (a) the cyclic mobility mechanism, the degradation of effective stress and large deformation under the undrained condition, and (b) shear strain accumulation and densification under the drained condition.     

基于深度学习的高分辨率遥感影像光伏用地提取

近年来我国光伏产业发展迅猛,随之也产生了诸多用地问题,通过遥感技术提取光伏用地,监测光伏用地分布与用地状况,对于光伏产业健康发展具有重要意义。本文提出一套基于深度学习方法的高分辨率遥感影像光伏用地自动提取方法,该方法利用GF-1等卫星影像和Google Earth影像构建光伏用地样本,基于ResNeSt-50作为骨干网络的DeepLab V3+模型实现深度学习语义分割算法,并结合计算机图形学方法对深度学习结果进行后处理,实现了面向高分辨率遥感影像较通用的且高精度的光伏用地自动提取。该方法的深度学习模型验证精度mIoU值达0.899 2,提取结果具有良好的边缘精度且具有广泛的适用性,支持GF-1、ZY-3、GF-6、GF-2和Google Earth等影像。     

时间序列SAR图像不可移动文物水域淹没监测

全球气候变化引起的极端降雨、洪涝灾害是导致不可移动文物受损的重要破坏因素。合成孔径雷达SAR（Synthetic Aperture Radar）具有全天时、全天候、大范围周期性对地观测的优势,是大区域水体监测的重要手段,对不可移动文物水域淹没及风险监测具有重要意义。本文利用时间序列SAR图像提出基于残差U-Net的不可移动文物水域淹没及风险监测框架。首先,基于双峰阈值分割法结合专家知识辅助进行水体样本生成,提高样本制作效率;其次,引入残差模块建立U型卷积网络,综合残差结构及U-Net的特点,缓解梯度更新时的弥散、消失等现象,通过卷积层之间累加和跳跃链接,保留多尺度的地物特征信息,以实现水体快速、高精度的语义分割;最后通过将结果与不可移动文物点位进行空间叠加分析,实现对不可移动文物水淹状况的监测。选取鄱阳湖及南昌市昌邑北垱遗址作为试验研究对象。获取了21景覆盖鄱阳湖区域不同时相Sentinel-1 SAR图像,并结合Sentinel-2光学图像进行结果分析与评价。实验结果表明：本文方法在鄱阳湖试验区对水体提取总精度优于95%,相较于FCN（Fully Convolutional Networks）与U-Net方法具有更好的精度。利用不同时相SAR图像获得时间序列水体分布范围变化图,与昌邑北垱遗址点位进行空间叠加分析,得到不可移动文物水域淹没长时间序列监测结果。实验结果表明本文提出的方法可以有效提取水体范围,对不可移动文物水域淹没及风险监测可以提供有力支撑。     

中国地质学的产生与发展——写在地质调查所创建100周年之际

中国的地质学非我国自有,而是在传播和引入西方近代科学文化的过程中逐渐形成的,是"西学东渐"的结果。在欧美等西方国家,至19世纪中叶,便建立并完善了近代地质学的理论和方法体系,完成了学科体制化建设,比中国至少早一个世纪。从16世纪后半叶的"西学东渐",西方近代地学传入中国进行第一次启蒙,到引进、消化、吸收和本土化,最后建立中国地质学理论和方法体系,完成学科建制,中国地质事业历经了长达三个半世纪艰辛曲折并充满坎坷的历史进程。创建于1913年的地质调查所是中国近代第一个科研机构,在章鸿钊、丁文江、翁文灏等第一代地质学家的领导下,筚路蓝缕、艰苦奋斗,在短短的二、三十年里,全面开创和奠定了中国近代地质事业,培养了一大批早期地质学家,取得了举世瞩目的成果,赢得国内外学术界的广泛赞誉。地质调查所的成功及其所取得的成就远远超出了地质学,对我国古生物学、地理学、地震学、地球物理学、土壤学、考古、古人类学、燃料和地图等诸多学科领域都产生了深远影响。     

Recent advances in earthquake monitoring II: Emergence of next-generation intelligent systems

Seismic data processing techniques, together with seismic instrumentation, determine our earthquake monitoring capability and the quality of resulting earthquake catalogs. This paper is intended to review the improvement of earthquake monitoring capability from the perspective of data processing. Over the past two decades, seismologists have made considerable advancements in seismic data processing, partly thanks to the significant development of computational power, signal processing, and machine learning techniques. In particular, wide application of template matching and increasing use of deep learning significantly enhance our capability to extract signals of small earthquakes from noisy data. Relative location techniques provide a critical tool to elucidate fault geometries and seismicity migration patterns at unprecedented resolution. These techniques are becoming standard, leading to emerging intelligent software systems for next-generation earthquake monitoring. Prospective improvements in future research must consider the urgent needs in highly generalizable detection algorithms (for both permanent and temporary deployments) and in emergency real-time monitoring of ongoing sequences (e.g., aftershock and induced seismicity sequences). We believe that the maturing of intelligent and high-resolution processing systems could transform traditional earthquake monitoring workflows and eventually liberate seismologists from laborious catalog construction tasks.     

Spectral sparse Bayesian learning reflectivity inversion

A spectral sparse Bayesian learning reflectivity inversion method, combining spectral reflectivity inversion with sparse Bayesian learning, is presented in this paper. The method retrieves a sparse reflectivity series by sequentially adding, deleting or re‐estimating hyper‐parameters, without pre‐setting the number of non‐zero reflectivity spikes. The spikes with the largest amplitude are usually the first to be resolved. The method is tested on a series of data sets, including synthetic data, physical modelling data and field data sets. The results show that the method can identify thin beds below tuning thickness and highlight stratigraphic boundaries. Moreover, the reflectivity series, which is inverted trace‐by‐trace, preserves the lateral continuity of layers.